- FugakuNEXT will mix Fujitsu CPUs and Nvidia GPUs in Japan’s subsequent nationwide supercomputer
- The system targets 600EFLOPS FP8 efficiency with a 100x utility efficiency achieve
- Riken, Fujitsu, and Nvidia see the mission as defining a brand new AI-HPC normal
Japan is making ready its subsequent nationwide supercomputer, FugakuNEXT, by way of a collaboration between Fujitsu, Nvidia and Riken.
The system is deliberate for operation round 2030 and goals to mix simulation and synthetic intelligence into one tightly built-in platform.
For the primary time in a Japanese flagship mission, GPUs will probably be used as accelerators. Nvidia will (unsurprisingly) design the GPU infrastructure, Fujitsu will deal with CPUs and system integration, and Riken will probably be concerned within the software program and algorithm work.
Feynman GPU
The result’s anticipated to be an “AI-HPC platform” designed for science, trade, and AI-driven discovery.
The efficiency targets for the supercomputer are definitely formidable. FugakuNEXT is designed to ship greater than 600EFLOPS of FP8 AI efficiency, which might make it probably the most highly effective AI supercomputer but introduced.
The system can be anticipated to realize as much as a hundredfold improve in utility efficiency in contrast with Fugaku, whereas staying inside roughly the identical 40MW energy funds.
Nvidia’s long-term roadmap factors to the Feynman GPU structure (named after theoretical physicist Richard Feynman) arriving close to 2028, so it might properly play a task in powering FugakuNEXT.
Fujitsu is growing a successor to its MONAKA CPU for the mission, tentatively named MONAKA-X, with extra cores, prolonged SIMD capabilities, and Arm’s matrix computation engine for AI inference.
Coupled with Nvidia’s accelerators, the system is anticipated to run giant simulations alongside demanding AI workloads.
{Hardware} alone received’t ship the goal positive aspects so the mission may even lean on improvements akin to surrogate fashions, mixed-precision arithmetic, and physics-informed neural networks to speed up efficiency whereas additionally preserving accuracy.
Makoto Gonokami, president of Riken, mentioned, “It’s a nice honor for Riken to collaborate with Fujitsu and Nvidia in advancing the event of FugakuNEXT. Since historic instances, humankind has constructed civilizations and superior societies by way of the science of computing. At the moment, the emergence of AI, superior semiconductors, and quantum computer systems is bringing a few discontinuous transformation in computational science.”
Ian Buck, vp at Nvidia, added, “FugakuNEXT will ship zettascale efficiency with utility speeds almost 100 instances sooner – throughout the identical vitality footprint as its predecessor – accelerating analysis, boosting industrial competitiveness, and driving progress for individuals in Japan and world wide.”
